Disentangled Latent Representations For Robust Speech Recognition: A Probabilistic Mel-Frequency Cepstral Coefficient–Variational Autoencoder Approach

Abstract

The performance of speech recognition systems greatly depends on how well we extract features from sound samples. Standard tools like MFCC (Mel-Frequency Cepstral Coefficient), PLP (Perceptual Linear Prediction), and LPC (Linear Predictive Coding) work well in quiet environments, but they often have trouble in the real world, where background noise distorts the sound signal. Even newer models, such as MFCC-GANs (Mel-Frequency Cepstral Coefficient - Generative Adversarial Network), sometimes fail to capture the complex and shifting nature of noisy settings. To address this issue, we suggest a hybrid feature extraction system that combines the reliability of MFCC with the flexibility of a VAE (Variational Autoencoder). The main idea is to use the VAE to map noisy speech data into a structured latent space. This allows the model to ignore random noise and focus on the key patterns of the human voice. Our results indicate that this hybrid approach is much more robust than traditional methods. In tests conducted under noisy conditions, the proposed model achieved the lowest Mean Squared Error (MSE) and surpassed MFCC, LPC, PLP, and GAN-based models. Additionally, a high silhouette score shows that our VAE-processed features are cleaner and easier for the system to differentiate. These findings suggest that using probabilistic models like VAEs can greatly enhance the reliability of speech recognition in everyday noisy environments.